Know Before Yes

Turn a website-and-app offer into a decision-ready draft—and export only the exact revision a human approves.

Comment

Problem

A nontechnical small-business owner can receive a polished website-and-app offer that names a price, timeline, and attractive outcomes while leaving ownership, data boundaries, the app's unique job, success measures, and long-term support unresolved. Those omissions are difficult to find and compare before the buyer commits to the wrong build.

Solution

Know Before Yes is a static WebMCP decision-planning tool from Sykes Holding Group LLC. Its default judge path uses a bundled fictional vendor offer. A secondary, human-activated lane accepts only closed, general, nonidentifying scenario categories.

An agent can discover seven focused tools, inspect the scenario, identify source-bound missing decisions, compare three practical paths, and stage a decision-ready draft. Premature commit and export calls return structured rejections without changing approved state. Only the visible human control can approve the exact staged revision. The agent can then read that approval status, commit the matching revision in the current session, and export the same revision as a local Markdown or JSON draft.

Agent analyzes. Human decides. The web tool exports what was approved.

Why This Matters

Small-business buyers often need structured technical judgment before they need a larger project-management system. Know Before Yes makes the missing decisions, tradeoffs, and authority boundary visible without asking the user to upload a proposal, create an account, or hand control to an agent.

WebMCP is central to the experience. The agent discovers typed capabilities and works through shared structured state instead of scraping cards, imitating clicks, or inventing facts. People and agents can now collaborate on a concrete draft while the page itself preserves the visible human decision boundary.

The current V1 demonstrates this workflow; it does not claim real-customer validation, vendor selection, purchasing authority, professional advice, guaranteed savings, guaranteed growth, or an external-system write.

How We Used AI

The released web tool runs no model and requires no API key. A WebMCP-capable browser or agent supplies reasoning outside the page and invokes explicit contracts. The page returns deterministic structured results and enforces staged-versus-approved state in its own local state machine.

During development, native WebMCP testing used OpenAI Codex's in-app browser and Google Chrome 151 with WebMCP enabled. The final permanent-host proof used OpenAI Codex's in-app browser. Two fresh native flows invoked all seven tools, returned both preapproval denials, recorded visible UI-only approval, read back human_approved and canCommit=true, and committed/exported exact revision rev-001-staged_platform. Clean run A completed in 42.564 seconds; judge-paced run B completed in 81.401 seconds. The final run produced a 2,153-character export and no browser-console error or warning. The clearly labeled Preview Adapter was used only for deterministic fallback rehearsal and is not presented as native proof.

How We Used Codex

OpenAI Codex and ChatGPT Work helped convert the initial product idea into a bounded scope, product requirements, technical specification, implementation checklist, static application, automated tests, browser-validation plan, public-release contract, and this Devpost draft.

Codex helped implement and challenge the seven-tool design, the exact-revision approval state machine, source-bound missing-decision results, structured preapproval errors, local export, lifecycle reset behavior, closed-enum scenario mode, privacy/claims boundaries, public release manifest, and automated product and viability gates. Independent technical, privacy/policy, and product-story reviews caught and corrected stale-session restoration, overstated build-ready wording, and nonfunctional structured inputs before the public candidate was frozen.

Approved AI-tool disclosure: OpenAI Codex and ChatGPT Work for product planning, implementation, debugging, testing, browser validation, evidence reconciliation, release preparation, and submission drafting.

Key Features

Structured offer inspection — returns scope, cost signals, exclusions, and clarity risks without changing decision state. Source-bound missing decisions — ties every question to a specific omission or general-category input instead of inventing evidence. Three practical paths — compares website-first, staged website-to-app, and combined-launch paths by fit, speed, cost shape, risk, and assumptions without claiming a universal best answer. Reversible staging — creates a deterministic revision id, leaves approved state unchanged, and invalidates stale approval. Visible human gate — no agent-facing approval tool exists; only the visible UI can record human-ui approval for the exact staged revision. Two explicit denials — premature commit returns HUMAN_APPROVAL_REQUIRED; premature export returns APPROVED_BLUEPRINT_REQUIRED. Exact-revision proof — approval status, committed state, UI, and Markdown/JSON export carry the same revision id. Two privacy-bounded modes — a resettable bundled fictional offer and an optional human-activated closed-enum general-category scenario. Static and state-free by design — no backend, accounts, persistence, analytics, uploads, payment, runtime dependency, or application network call.

WebMCP Tool Inventory

inspect_vendor_offer find_missing_decisions compare_solution_paths stage_blueprint_revision get_approval_status commit_approved_blueprint export_build_blueprint

The page defines exactly these seven tool contracts and conditionally registers them when document.modelContext is available. Structured scenario activation and human approval remain visible human UI actions, not eighth and ninth tools.

Architecture

Dependency-free static HTML, CSS, and browser JavaScript modules Pure in-memory decision domain in src/domain.js Seven WebMCP contracts and conditional registration in src/tools.js Shared UI/tool rendering, lifecycle reset, and user-initiated local download in src/app.js Static hosting headers for CSP, tools permission, origin isolation, MIME sniffing, referrer handling, framing/form restrictions, and HTTPS transport Node's built-in test runner plus focused P0 and viability gates No backend, database, login, analytics, telemetry, payment, OCR, upload, external model call, or persistence

The WebMCP client may process selected categories and tool results under its own terms and privacy settings. Static hosting may process ordinary connection and security metadata, but SHG does not operate a scenario-content endpoint.

Testing Instructions

Public judge flow — permanent host and final native proof verified

No credentials are required.

Open https://knowbeforeyes.sykesholdinggroup.com/ in ChatGPT's in-app browser or Google Chrome with WebMCP enabled. Select Reset and confirm the page shows the bundled fictional scenario. Discover these seven tools: inspect_vendor_offer, find_missing_decisions, compare_solution_paths, stage_blueprint_revision, get_approval_status, commit_approved_blueprint, and export_build_blueprint. Invoke the first three tools and confirm one source-bound gap and the three named path tradeoffs. Invoke stage_blueprint_revision for staged_platform; confirm approved state remains unchanged. Invoke commit_approved_blueprint; confirm HUMAN_APPROVAL_REQUIRED and approvedStateChanged=false. Invoke export_build_blueprint; confirm APPROVED_BLUEPRINT_REQUIRED. Use the visible Approve exact revision control. Invoke get_approval_status; confirm the staged revision id, source human-ui, status human_approved, and canCommit=true. Invoke commit_approved_blueprint, then export_build_blueprint; confirm staged, approved, committed, and exported revision ids match. Confirm the before/after summary, unresolved decisions, and local download boundary; then Reset again to prove repeatability.

This text is ready to paste into Devpost field 28255; the stable hostname and final signed-out run are verified. Update it only if the final evidence changes.

Repository verification

From the public repository root: npm test npm run p0:product npm run gate npm run serve

Open http://localhost:4173. Native discovery requires a supporting WebMCP client; otherwise the page truthfully identifies the Preview Adapter.

Judging-Criteria Map

WebMCP Leverage

Seven typed, non-overlapping contracts divide inspection, gap discovery, comparison, staging, status, gated commit, and export. The same structured state powers tool results, UI, rejection semantics, approval, and exact export. WebMCP is the operating collaboration layer, not a decorative prompt wrapper.

Execution

The product has a coherent resettable workflow, responsive UI, explicit errors, visible human decision control, local exports, a bounded second scenario mode, public source, and repeatable automated gates. The current exact public candidate passes 23/23 full tests, 10/10 P0 product checks, and 29/29 viability checks.

Potential Impact

Know Before Yes addresses a specific problem for nontechnical small-business buyers: polished proposals often hide the decisions that determine ownership, cost, scope, data boundaries, and long-term maintainability. The default deterministic scenario makes the judge proof repeatable, while the optional closed-enum mode demonstrates a reusable product shape without collecting proposals or customer data.

Creativity & Ambition

The product treats WebMCP as an authority-aware collaboration layer. An agent can prepare valuable work, but it cannot manufacture the human decision that makes a revision approved. The exact-revision commit/export proof turns human-in-the-loop language into visible, testable product behavior.

Known Limitations

One deterministic bundled fictional offer plus one optional closed-enum structured general-category mode No free text, URLs, uploads, documents, customer data, backend, account, collaboration, analytics, or persistence In-memory state only; startup, Reset, and every page restoration clear the session Approval is a visible workflow boundary, not identity verification, authentication, a legal signature, or an external security boundary No model runs inside the released page; WebMCP or AI-client processing is governed by that provider No real-vendor ranking, customer validation, purchasing, professional advice, savings claim, or outcome guarantee The public YouTube demo is hosted on the Sykes Holding Group channel and remains subject to YouTube availability and playback behavior

Built With

Share this project:

Updates